Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!swrinde!sdd.hp.com!zaphod.mps.ohio-state.edu!cis.ohio-state.edu!ucbvax!INEL.GOV!djy
From: djy@INEL.GOV (Daniel J Yurman)
Newsgroups: comp.society.futures
Subject: (none)
Message-ID: <9106251545.AA23629@gemstone.inel.gov>
Date: 25 Jun 91 15:45:44 GMT
Sender: daemon@ucbvax.BERKELEY.EDU
Organization: The Internet
Lines: 142

To: info-futures@world.std.com
Subject: Future of Super Computers

                                    Trip Report
Conference title:         Supercomputing USA / Pacific 91
                          Santa Clara, CA June 19-21, 1991

6/20/91

Steve Colley, N-Cube Corp.

-       The demand is for practical super computing.
-       Shift from research to production focus
-       Super computing is still technology driven, e.g. vlsi
-       Production issues will be
        *     price / performance
        *     reliability
        *     support
-       Future of corporations will be information intensive

Ed Masi, Cray Research

-       Connecting workstations to super computers is essential. 
        End-user interfaces are needed.  RISC-based workstations are
        linked to massively parallel architectures.  The workstation
        handles the visual interface to data.  The super computer
        handles the numerics.

-       1/2 of Cray's R&D $s are for software, especially
        application software which must be delivered simultaneously
        on super computers and workstations.

-       Parallel architectures are the backbone of high performance
        computing.  Good application software will be essential to
        success in this area.

-       Regarding I/O bottlenecks in massively parallel
        architectures (MPP), each processor will have access to I/O. 
        The design is for modest pipes, but lots of them.  The other
        I/O bandwidth bottleneck is between the workstation and the
        MPP.  Fiber optics can address this problem. 

-       With regard to the metrics of using arrays of micro
        processors, e.g. Intel i860, a limiting factor is the
        switching between multiple processors and memory.

-       Standards for programming languages will be C, Fortran 90,
        but most of all Fortran 77 due to the "billions" of line of
        code in the installed base.

-       MPP is not an "augmentation" technology.  It is a
        "replacement" technology and will be widely installed by
        93/94.  It will open up new classes of problem solving.

-       Japanese are better at exploiting MPP than US.  Nissan has
        more computing power than all US automakers combined.


Bob Palluck, Convex Corp.

-       Parallel architectures are allowing oil & gas firms to shift
        from 2D to 3d visualization of seismic data.

-       In mechanical engineering new technologies will allow the
        engineer to spend more time on the design problem and less
        on specification of computer code.  In computational
        chemistry simulation of proteins will be speeded up.  In
        weather forecasting, grid spacing is decreasing.  Also,
        models now take terrain variations into account.

-       Simulators will allow the engineer to get design
        alternatives back more quickly.  

-       There will be a two-tier environment.  It will link the
        workstation with shared super computer resources which are
        10x faster than the workstation.  User organizations will be
        dealing with heterogeneous standards.

-       Super computers will be characterized by high-speed
        backplanes, average memory of 250 Mb to 1 Gb, shift from 32-
        bit to 64-bit data paths, and utilization of workstation
        operating system features such as those found in MACH.  We
        are looking at 2-year technology refreshment cycles.

-       There is a massive hole in the market in terms of a lack of
        application software to take advantage of "miller micro"
        workstations linked to MPP.


6/21

Jim Clark, Silicon Graphics

-       MPP will support new multi-media applications including
        HDTV.  1 Gigaflop / sec compute power on a chip will support
        compression / decompression of graphic images in
        workstations and eventually home consumer electronics.

Randy Groves, IBM

-       Workstation trends focus on;
        *     price / performance
        *     processor
        *     memory
        *     connectivity

-       Optical communications will result in greater bandwidth,
        greater overall mips, support for multi-media and virtual
        reality.  Performance of 100 Mb / sec is needed.


Bill Poduska, Stardent

-       Issues are technology, markets, and local v. central
        computation.

-       Major technology trend is use of database technology to
        retrieve visual images.  It is not clear that object
        oriented DBMS is the model for this work.  Research is
        needed on database structures to support visualization and
        multi-media.

-       Performance issues will be volume of information retrieved,
        calculations made, and synthesis of data into images. 
        Ultimately, to move MPP into production, firms will ask
        about return on investment while faced with rapid technology
        refreshment cycles.

----------------------------------------------------------------
Notes prepared by

   ...!standard!disclaimer@INEL!included
/* -------------------------------------------------------- */
   Dan Yurman                      Internet: djy@inel.gov
   Idaho National Engineering Lab.      Map: 43N 112W -6GMT
   PO Box 1625  MS 3900               Phone: (208) 526-8591
   Idaho Falls, ID 83415                Fax: (208) 524-6374
/* -------------------------------------------------------- */
   In a process in which requirements are not set or vague,
   nonconformance and quality will be randomly distributed!

PS:	I welcome any corrections or additions to these notes.